[R-sig-ME] ML or REML for LR tests

[Austin Frank]

First off, thanks to all who have responded to the series of questions I
asked!

On Fri, Aug 29 2008, Ken Beath wrote:

> On 29/08/2008, at 2:47 PM, Austin Frank wrote:
>
>> 3) Is it the case that LR tests between REML models with different
>> random effects are meaningful?  Does this apply to both nested and
>> non-nested models?
>>
>
> Maybe, but only for nested (see Q2). Supposedly it works better than
> ML. The significance tests wont be correct but if there is a huge
> significance level then there is probably a random effect. Simulation
> seems a better idea.

Ken was the only one to address this particular point, and I want to
make sure I've got it straight.  Are REML-based likelihood-ratio tests
(presumably not performed with anova.mer, as that sets REML=FALSE on the
call to logLik) an acceptable method for testing nested models with
different random effects specifications?

As a point of reference, the anova() method is called on two lmer models
that differ only in their random effects in the manuscript by Baayen,
Davidson, and Bates at
http://www.ualberta.ca/~baayen/publications/baayenDavidsonBates.pdf (pp
12-15).  The discussion of that analysis makes no mention of the
difference between REML and ML fits.  Is this because, as discussed
recently, the REML and ML estimates are so close that there is no
practical difference in which quantity is used for this test?


Thanks again!
/au

-- 
Austin Frank
http://aufrank.net
GPG Public Key (D7398C2F): http://aufrank.net/personal.asc